Source-linked AI summary
Rice Diseases Detection and Classification Using Attention Based Neural Network and Bayesian Optimization
Yibin Wang, Haifeng Wang, Zhaohua Peng
TL;DR
Rice disease affects rice growth and development, while plant diseases contribute to substantial crop losses and economic costs. The paper proposes an attention-enhanced DNN model and reports that its ADSNN-OB model outperforms tested deep-learning and state-of-the-art models in classification.
Problem
Rice disease affects rice growth and development, while plant diseases are associated with 20 to 40 percent of global crop production losses and substantial economic costs.
Method
The paper proposes an attention mechanism to enhance DNN model learning and introduces a model named ADSNN-OB, using pre-trained ImageNet weights for better results.
Results
The proposed ADSNN-OB model outperforms other deep-learning models and all tested state-of-the-art models in classification.
Takeaways & Limitations
The study supports the proposed model as a strong candidate for rice disease classification across the tested scenarios.
Takeaways & Limitations
The study notes that its dataset differs from those used in earlier studies and plans further exploration of the proposed model.
Abstract
from arXiv · showhide
In this research, an attention-based depthwise separable neural network with Bayesian optimization (ADSNN-BO) is proposed to detect and classify rice disease from rice leaf images. Rice diseases frequently result in 20 to 40 \% corp production loss in yield and is highly related to the global economy. Rapid disease identification is critical to plan treatment promptly and reduce the corp losses. Rice disease diagnosis is still mainly performed manually. To achieve AI assisted rapid and accurate disease detection, we proposed the ADSNN-BO model based on MobileNet structure and augmented attention mechanism. Moreover, Bayesian optimization method is applied to tune hyper-parameters of the model. Cross-validated classification experiments are conducted based on a public rice disease dataset with four categories in total. The experimental results demonstrate that our mobile compatible ADSNN-BO model achieves a test accuracy of 94.65\%, which outperforms all of the state-of-the-art models tested. To check the interpretability of our proposed model, feature analysis including activation map and filters visualization approach are also conducted. Results show that our proposed attention-based mechanism can more effectively guide the ADSNN-BO model to learn informative features. The outcome of this research will promote the implementation of artificial intelligence for fast plant disease diagnosis and control in the agricultural field.
1. Introduction
Rice diseases threaten crop production and require timely diagnosis, yet rice disease diagnosis remains manual and labor-intensive. The paper proposes an attention-based, mobile-compatible CNN with Bayesian optimization and evaluates it against other deep-learning models.
- Motivation: 20 to 40 percent of global crop production is lost to diseases and pests, with plant diseases costing around $220 billion annually.
- Motivation: Unrestrained pesticide use can cause considerable environmental damage, making precise and timely disease diagnosis critical.
- Motivation: Rice disease diagnosis is often manual, requiring special skills and equipment that can delay treatment or cause misdiagnosis.
- Proposed approach: The study develops an attention mechanism to guide deep networks toward critical and informative features for rice disease classification.
- Proposed approach: The proposed CNN applies attention-based depthwise separable convolution and is designed for portable devices such as mobile phones and single-board computers.
- Evaluation: ADSNN-BO outperforms all tested state-of-the-art models in rice disease classification accuracy and other performance measurements.
2. Literature Review
Prior rice-disease studies applied machine learning, deep learning, transfer learning, IoT, and image-processing methods, while highlighting accuracy, mobile deployment, and interpretability considerations. These works motivate a mobile-friendly model with stronger feature analysis and optimization.
- Multiple studies applied deep learning and machine learning to rice-disease detection and classification.
- 93.3% accuracy was achieved by a two-stage small CNN architecture designed with substantially reduced model size for mobile-device suitability.
- CNN features were more discriminative and representative than LBPH and Haar-WT, while CNN and CNN-SVM performed similarly and exceeded LBP-SVM and Haar-WT-SVM.
- 98.9% accuracy was reported for optimized DNN classification of healthy, bacterial blight, brown spot, sheath rot, and blast leaf images.
- The proposed research targets an optimized mobile-friendly model and adds intermediate activation, convolutional-layer feature-pattern, and useful-filter visualization.
3. Attention-based Depthwise Separable Neural Network with Bayesian Optimization
The proposed ADSNN-BO model combines MobileNet’s depthwise separable convolutions with augmented attention and Bayesian optimization to support efficient rice-disease image classification.
- Model overview: ADSNN-BO combines a MobileNet-based architecture with an augmented attention mechanism and Bayesian optimization.The model is designed to address accuracy, size, latency, and hyperparameter-selection considerations.
- Bayesian optimization: The study applies Bayesian optimization to tune model hyperparameters and explore the number of attention layers.Bayesian optimization uses a statistical model of the objective function and an acquisition function to select evaluation points.
- Depthwise separable architecture: MobileNet uses depthwise separable convolutions, which separate channel-wise filtering from cross-channel combination.Depthwise convolution applies one filter per input channel, while pointwise convolution uses a 1 × 1 operation to combine outputs.
- Depthwise separable architecture: Depthwise separable convolution reduces model size and computational complexity relative to standard convolution.The operation decomposes conventional convolution into filtering and combining stages, reducing the interaction between kernel size and output channels.
- Attention integration: Attention layers are inserted between the last convolution block and average pooling while retaining the original MobileNet structure and hyperparameter settings.The detailed architecture is presented in Figure 1, with convolution blocks and their outputs indicated.
- Feature analysis: Activation maps, filters, and heatmaps are used to inspect learned features and evaluate how the convolutional model processes images.Filter visualization maximizes a selected filter’s response through gradient ascent in input space.
4. Experiments
Experiments evaluate ADSNN-BO for classifying rice-leaf diseases using cross-validated comparisons with established deep-learning models. Results show stronger overall and class-level performance, while feature visualizations indicate disease-relevant activations and less noisy filters.
- Data Preparation: The study focuses on brown spot, rice hispa, rice leaf blast, and healthy rice classes using a dataset of 2370 rice-leaf samples.
- Performance Measurement: Models are assessed with five-fold cross-validation using confusion-matrix-derived precision, recall, and F1-score measurements.The reported table values are summarized as means and standard deviations across the five folds.
- Performance Measurement: ADSNN-BO achieves the highest test accuracy, precision, and F-1 score among the tested models, while original MobileNet achieves the highest recall.ADSNN-BO’s training time is close to that of original MobileNet.
- Performance Measurement: ADSNN-BO improves precision by 8.18%, F-1 score by 3.23%, and accuracy by 3.6% compared with original MobileNet.
- Interpretability Analysis: ADSNN-BO outperforms other deep-learning models for each rice-disease category, while leaf blast is the most difficult disease to detect and classify.Activation maps show disease-specific patterns, including spot patterns for brown spot disease, and ADSNN-BO filters are more recognizable and less noisy than original MobileNet filters.
5. Conclusions
The study compares deep learning models for rice disease classification and proposes ADSNN-OB as a mobile-compatible model that outperforms the tested alternatives. Feature visualization supports analysis of model performance, while future work targets broader optimization, datasets, and contextual information.
- Pre-trained ImageNet weights were used across the CNN architectures, alongside data preparation and varied classification experiments.
- ADSNN-OB was proposed, tested, and discussed as a new model for rice disease classification.
- The proposed ADSNN-OB model outperforms other deep learning models across various scenarios and tested state-of-the-art comparisons.
- Activation-map and filter visualizations were implemented to evaluate model features and performance.
- ADSNN-OB fits the mobile-device environment, while future work will examine other optimization methods, hyperparameter tuning, public datasets, and location, weather, or soil information.